Artificial Synapses Based on Multiterminal Memtransistors for Neuromorphic Application

Abstract

Neuromorphic computing, which emulates the biological neural systems could overcome the high‐power consumption issue of conventional von‐Neumann computing. State‐of‐the‐art artificial synapses made of two‐terminal memristors, however, show variability in filament formation and limited capacity due to their inherent single presynaptic input design. Here, a memtransistor‐based artificial synapse is realized by integrating a memristor and selector transistor into a multiterminal device using monolayer polycrys‐talline‐MoS2 grown by a scalable chemical vapor deposition (CVD) process. Notably, the memtransistor offers both drain‐ and gate‐tunable nonvolatile memory functions, which efficiently emulates the long‐term potentiation/depression, spike‐amplitude, and spike‐timing‐dependent plasticity of biological synapses. Moreover, the gate tunability function that is not achievable in two‐terminal memristors, enables significant bipolar resistive states switching up to four orders‐of‐magnitude and high cycling endurance. First‐principles calculations reveal a new resistive switching mechanism driven by the diffusion of double sulfur vacancy perpendicular to the MoS2 grain boundary, leading to a conducting switching path without the need for a filament forming process. The seamless integration of multiterminal memtransistors may offer another degree‐of‐freedom to tune the synaptic plasticity by a third gate terminal for enabling complex neuromorphic learning.

Spiking neurons from tunable Gaussian heterojunction transistors

“…Not only is the fundamental behavior of the Na + ion channel of a biological neuron captured by the GHeT in a simple circuit, but by exploiting the dual-gated programmability both through independent and dependent biasing, it is possible to achieve eight different biological neuron responses, five of which are achieved using a single GHeT, two transistors, two capacitors, and two resistors. Additionally, the fabrication process for GHeT-based spiking neurons is compatible with previous demonstrations of monolayer MoS 2 memtransistor-based synapses 12,13 , enabling scalable implementations of biomimetic neuromorphic platforms. More broadly, since CMOS transistors cannot natively mimic the Gaussian response demonstrated here, CMOS-based digital designs implement Gaussian functions with complex circuits and look-up tables while analog CMOS circuits suffer from limited programmability and high bias current 47 .…”

“…To address the limitations of silicon-based SNN circuits, alternative materials are being explored that allow the encoding of neuromorphic functionality directly at the device level. While memristors 11 , memtransistors 12,13 , domain-wall memories 14 , metal-insulator-transition (MIT) devices 15 , multi-gated transistors 16,17 , and Gaussian synapses 18 have been developed for scalable implementation of synaptic functions, approaches for realizing spiking neurons are relatively lacking. For example, neuristors based on MIT devices have been reported, but this design suffers from low gain and limited output swing 19,20 .…”

Bridging Biological and Artificial Neural Networks with Emerging Neuromorphic Devices: Fundamentals, Progress, and Challenges

“…[123,[140][141][142] Also, FETs based on silicon, organic materials, perovskite or carbon nanotube (CNT) have been reported as synaptic elements based on either charge trapping/detrapping mechanism or floating-gate (FG) memory structure, some of which have a structure of multi-gate FET. [120][121][122][123][124][125][126][127][128] Three-terminal FeFET and two-terminal ferroelectric tunnel junction (FTJ) relying on ferroelectric polarization switching have been explored as artificial synapses, where the pulsing scheme needs to be carefully designed to improve the switching symmetry and linearity. [94,103] Similarly, two-terminal spin-transfer torque MRAM (STT-MRAM) based on magnetization switching and multi-terminal magnetic devices based on domain wall motion have also been studied to implement artificial synapses.…”

Two-Dimensional Near-Atom-Thickness Materials for Emerging Neuromorphic Devices and Applications

“…Gate-tunable memristors, or memtransistors, are realized by combining the concepts of both transistor and memristor into a single device. Memtransistors offer both drain- and gate-tunable NVM functions, which efficiently emulate the long-term potentiation (LTP)/depression (LTD), spike-amplitude, and spike-timing-dependent plasticity (STDP) of biological synapses ( Wang et al., 2019a ). Laterally configured 2D materials are beneficial for devising 2D vdW heterostructures-based memtransistors by vertically assembling 2D layers in an open architecture.…”

“…Laterally configured 2D materials are beneficial for devising 2D vdW heterostructures-based memtransistors by vertically assembling 2D layers in an open architecture. Many studies employed CVD-grown monolayer MoS 2 in lateral memtransistors ( Jadwiszczak et al., 2019 ; Sangwan et al., 2015 , 2018 ; Wang et al., 2019a ; Xie et al., 2017 ). Sangwan et al.…”